Insights lesenGespräch

Judgment

The bottleneck shifts to judgment.

Drafting Gets Cheap. Judgment Gets Valuable.

That AI speeds up work is well known. What matters more is what shifts as a result: when output costs almost no time, judging it becomes the bottleneck. This affects the whole organization, all the way up to the executive team.

11 min read

Who signs all this off?

Contents

Most conversations about AI revolve around one question: what can it do already? Texts, concepts, analyses, images, code. The list gets longer every month, and the results keep improving.

That AI speeds up work is by now well known. What matters more is the shift this creates. For a long time, producing the work was the time-consuming part. Now judging it becomes the decisive part.

That raises a different question: Who signs all this off?

Every AI output that is meant to be used further needs an accountable assessment. Depending on the risk, a standardized check may be enough, but often a human expert has to review it. This work does not get faster just because drafting has. It becomes the new AI productivity bottleneck, and not only for the quality of individual outputs but for the organization as a whole.

Drafting gets cheap

The effect of AI on drafting is well documented. In an MIT experiment with 453 professionals, participants using ChatGPT took about 40 % less time on typical writing tasks from their occupation. Independent professionals rated the quality of the results 18 % higher. The tasks were simplified and did not require detailed factual knowledge about a real company. The direction is still clear.

Since then, the tools have improved considerably. A first version, a variant, a summary or an analysis now costs almost nothing. Someone who used to spend three days on a concept now has a usable draft after an hour. Someone who used to present one variant can now present five.

That sounds like a pure gain. But it is only half the calculation.

The bottleneck moves from drafting to review

Four ideas lie behind this shift.

Where the bottleneck forms

Production scalesDrafts, analyses, images and code in minutes
Judgment does notA person checks facts, context and fit
Decisions pile upWho approves? Who is accountable?
Adapt how work is doneSet clear rules for review, approval and accountability

Production scales. Texts, analyses, concepts, images and code are produced in a fraction of the time, along with any number of variants. Whoever wants more gets more.

Human judgment does not scale automatically. Errors can hide anywhere: a wrong figure in the text, a skewed assumption in the analysis, an unsuitable detail in the image, a gap in the concept, a silent bug in the code. Someone has to assess facts, context, quality and fit. That takes knowledge of the company, its customers and the market. This work cannot be multiplied at will.

Decisions become the bottleneck. Who is allowed to approve an output? Who carries the responsibility? And who recognizes when an output looks good but is wrong?

The way of working has to change. Not only creating needs clear processes, but also review, approval and accountability.

Every organization has a limited capacity to review and decide. It usually rests with a few people: executives, department heads, experienced specialists. Until now, this limit was barely noticeable because drafting itself was slow. The backlog formed earlier in the process.

With AI, the backlog moves. When a team delivers twice as many drafts, twice as many land on the desks of those who approve. If that capacity stays the same, one of three things happens:

  • Work piles up. Drafts sit, projects wait for a decision, and the speed gained in drafting evaporates.
  • Review becomes superficial. Approvals get waved through because time is short. Errors that look convincing slip past.
  • Work is done twice. Unchecked outputs move on and have to be corrected later.

That leads to the real point. A company can multiply its output with AI and get slower at the same time, because its capacity to decide does not grow with it. The tools are not the problem.

What the numbers show

In software development, this can already be measured, because every change is reviewed before it is merged. The analytics company Faros AI evaluated data from more than 10,000 developers across 1,255 teams. The result fits the pattern described above.

+98 %more merged changes (pull requests) in teams with high AI useFaros AI, 2025
+91 %longer review time for these changesFaros AI, 2025
≈ 40 %of the time saved by AI is lost to rework: correcting, rewriting, checkingWorkday, 2026

More work, but not done faster

Merged changes: +98 %
Review time per change: +91 %
Tasks completed: +21 %
Bugs per developer: +9 %
Teams with high AI adoption compared with low-adoption teams. Faros AI 2025, more than 10,000 developers across 1,255 teams. No measurable improvement showed up at the company level.

The teams completed more tasks and delivered almost twice as many changes. But review time rose by almost the same amount, the changes got larger, and the number of bugs per developer increased slightly. At the company level, Faros found no clear link between AI use and better outcomes. The gain in drafting did not reach the organization. The data comes from software development and cannot be transferred one to one. It shows a correlation, not conclusive proof that AI alone causes the longer review times. But it shows where the bottleneck moves when drafting gets faster.

Outside of engineering, a Workday survey paints a similar picture. It covered 3,200 employees who actively use AI, at companies with at least $100 million in annual revenue. Most save between one and seven hours per week with AI. But almost 40 % of that time is lost again because outputs have to be corrected, rewritten and checked. Only 14 % of respondents say they consistently achieve a clear net gain.

How companies spot the bottleneck

In everyday work, the new bottleneck rarely shows up as a number. It shows up more in sentences like these:

  • “It’s still with management.”
  • “I made three versions, which one do we go with?”
  • “The AI wrote that, I only skimmed it.”
  • “We’re still waiting for approval.”

The more AI a team uses, the more often such sentences come up. This is not a failure of the employees. It is a sign that the organization was built for slow drafting, not fast drafting.

Where the time goes

Before
  • Drafting takes a long time
  • Few drafts, well prepared
  • Approval keeps up because little arrives
  • The bottleneck is in making
With AI
  • Drafting takes minutes
  • Many drafts and variants
  • Approval piles up or turns superficial
  • The bottleneck is in judging

All the way up to the executive team

The bottleneck is not limited to individual departments. It reaches the executive team, and it hits especially hard there.

AI delivers a market analysis, a summary of customer feedback or an evaluation of sales figures within minutes. The result looks finished: clear structure, clean language, an unambiguous conclusion. That is exactly what makes it risky. A quickly generated analysis is easily read as fact, even though it rests on incomplete data or a wrong assumption.

In the past, an analysis that reached the executive team had usually passed through several hands. With AI, it can come straight from a question, without anyone having checked the foundation. Anyone who derives decisions from it, on product range, budget or staffing, makes them faster. Whether they are better depends on how well the foundation was assessed.

For leaders, this means their own judgment and the question “What is this based on?” become more important, not less.

Equally polished, not equally good

AI creates a new problem that many teams already feel. Outputs that visibly come from AI initially earn less trust. After all, how deep the work behind them goes is hard to tell from the outside.

Two cases look identical on the desk. In the first, someone works thoroughly, knows the subject and uses AI to express their expertise faster and more clearly. In the second, someone uses the same tools, and the result looks just as convincing. But there is little underneath.

In the past, the difference was often visible. Thorough work looked thorough. That signal is gone. This has two consequences: reviewers can no longer rely on first impressions. And good work gets devalued along with the rest, because it can no longer be told apart from the outside.

Same on the outside, different underneath

Deep work, sharpened with AI
  • Looks professional
  • Sources and assumptions are known
  • Discarded options can be explained
  • Holds up to the second question
Generated with AI, little behind it
  • Looks just as professional
  • Sources and assumptions stay unclear
  • Reasons are missing
  • Falters at the second question

Trust therefore shifts from the document to the conversation and the person. A few simple habits help:

  • Ask follow-up questions. Why this way? Where does this number come from? What was discarded? Someone who has mastered a topic can answer. Someone who has only generated a good-looking result will show it by the second question.
  • Provide the foundation. Every output should come with its sources, its assumptions and what was deliberately left out. That makes the depth visible without recalculating everything.
  • Separate the certain from the uncertain. Proven, assumed, estimated: whoever makes that distinction earns trust. Whatever always sounds equally convincing rightly deserves less of it.
  • Let a track record count. As with any new team member, trust grows over time. Those whose outputs have proven reliable need less review.

In the end, a simple rule applies: whoever approves an output takes responsibility for that approval, regardless of whether and how AI contributed. The question is then no longer whether the AI can be trusted, but whether the person who approved it can. Companies have always known that question.

What helps: quality control for AI output

The answer is not to use less AI. The answer is to take reviewing and deciding as seriously as drafting.

Set criteria before the assignment. Whoever says up front what a good result looks like reviews faster afterward. A draft without a clear goal can only be judged with a lot of effort.

Tier approvals by risk. Not every AI output needs the same review. What matters is a proportionate, risk-based approval process. Regulating this clearly relieves the bottleneck.

Review by risk

Low
  • Internal idea draft, outline, summary for oneself
  • A quick look is enough
  • Approval by the person working with it
Medium
  • Ordinary product copy, customer emails, internal analyses for discussion
  • Check facts and figures, assess tone and fit
  • Approval by the subject-matter owner
High
  • Material price changes, health- or legally relevant product claims, automated decisions with significant impact, foundations for major investments
  • Thorough review, sources, and when in doubt, specialists or legal counsel
  • Approval with clear accountability, documented

The classification is illustrative. What matters are the content, the intended use and the possible consequences of an error.

Move decisions closer to the experts. Whoever can assess the work on its merits should also be allowed to approve it. This takes trust and clear limits. But it is the most effective lever against the backlog.

Fewer variants, better variants. AI makes it easy to deliver five versions. Each one costs review time. Often a well-reasoned recommendation is worth more than a selection.

Plan review capacity. When starting an AI project, don’t ask only how much faster drafting will get. Also ask: who reviews this, and does that person have the time? A project that only speeds up creation just moves the backlog.

This fits what the article AI Is a Tool already says for the individual: being good means being able to judge the result. For a company, that means being good at organizing judgment and decision authority where they are needed.

Judgment gets valuable

This is more than a question of productivity. It changes where value is created in a company. It used to lie largely in producing: whoever could draft faster and better had the edge. When drafting gets cheap, value shifts to judging.

What is needed are people who can quickly size up a draft. People who see what is missing, what is wrong and what moves the company forward.

That takes judgment and a feel for context. AI can deliver ten variants. Which one holds up is decided by a person who knows the audience, the company’s culture and the limits that must not be crossed.

And it takes accountability. A machine can make suggestions. It cannot answer for the consequences. That responsibility cannot be handed to a tool.

Then there is competition. When more and more companies use similar AI tools, the sheer ability to produce output is hardly an advantage anymore. What counts is who can better sort, check and turn results into good decisions.

AI lowers the effort for the first draft. It does not make experience, expertise and a sense of responsibility obsolete. On the contrary: the more outputs there are, the more important it becomes to tell the difference between usable, correct and truly good.

For leaders of mid-sized companies, this means a new task: not just introducing tools, but shaping review, approval and accountability so that the organization keeps pace.

Drafting gets cheap. Judgment gets valuable.

Key takeaways

  • AI makes drafting fast and cheap: drafts, concepts, analyses and code take minutes instead of days.
  • Reviewing, deciding and taking responsibility don't get faster as a result. They become the bottleneck.
  • This is already measurable: teams with heavy AI use deliver much more work, but review time rises sharply, and little reaches the company level.
  • A company can multiply its output with AI and get slower at the same time, because its capacity to decide does not grow with it.
  • Not every output needs the same level of review. An internal idea draft is different from a product claim, a customer email or an automated pricing decision. Approvals should be tiered by risk.
  • Thorough and superficial work look equally good with AI. Trust is therefore built in conversation: through follow-up questions, visible foundations and a track record over time.
  • This holds all the way up to the executive team: quickly generated analyses look finished and are easily misread.
  • When everyone uses similar tools, producing output is no longer an advantage. The difference is made in sorting, checking and deciding: between usable, correct and truly good.

Frequently asked questions

Does this mean AI brings no productivity gain?

No. In drafting, the gain is clear and well documented. But it only reaches the company if review and approval keep pace. Otherwise it evaporates in backlogs or rework.

Can’t AI take over the reviewing, too?

Partly. AI can flag errors, cross-check facts or compare drafts against criteria. But the decision and the responsibility stay with a human. This applies especially to statements made to customers, and to anything involving money or legal matters.

Where is the best place for a company to start?

With the question of where drafts are already waiting for approval today. That is where clear criteria and tiering by risk pay off first. Often a single page of rules on who may approve what is enough.

How can you tell whether real work stands behind an AI-assisted result?

Often no longer from the result itself. Better are follow-up questions about sources, assumptions and discarded alternatives. Whoever has mastered the topic can answer them. Added to that is experience of whose outputs have proven reliable.

Does this apply to the executive team as well?

Yes, especially there. Quickly generated analyses and summaries look finished and are easily read as fact. The question “What is this based on?” therefore belongs to every decision built on an AI analysis.

Does this affect only large companies?

In mid-sized companies, the effect tends to be stronger. Decisions there often rest with very few people. When those people become the bottleneck, the whole company feels it.

Sources

Inhalt

    InsightsIns Gespräch kommen